Slashdot Mirror
Why Do Computers Still Crash?
geoff lane asks: "I've used computers for about 30 years and over that time their hardware reliability has improved (but not that much), but their software reliability has remained largely unchanged. Sometimes a company gets it right -- my Psion 3a has never crashed despite being switched on and in use for over five years, but my shiny new Zaurus crashed within a month of purchase (a hard reset losing all data was required to get it running again). Of course, there's no need to mention Microsoft's inability to create a stable system. So, why are modern operating systems still unable to deal with and recover from problems? Is the need for speed preventing the use of reliable software design techniques? Or is modern software just so complex that there is always another unexpected interaction that's not understood and not planned for? Are we using the wrong tools (such as C) which do not provide the facilities necessary to write safe software?" If we were to make computer crashes a thing of the past, what would we have to do, both in our software and in our operating systems, to make this come to pass?
Even with my uptime experiments, which consisted of taking an old but reliable hardware, installing Windows 95/OSR2/98/98SE/ME, and then letting the computer idle and do nothing never resulted in more than about 25 days before I came over and windows was fubar'ed or the computer was simply locked hard.
Windows 3.1 actually did quite well if I remember right, as it seemed perfectly content sitting idle doing nothing seemly forever. Windows 9x always seemed to randomly thrash the HDD, even after a clean install, which led me to believe that Windows 9x is never truly idle, it's always up to something (virtual memory?), and that something eventually will bring it down.
Windows 9x actually has a bug in it that would lock the computer after 46 days of uptime, but it took years to catch it because no one ever got close to that mark.
A lot of people are answering the question of why there are bugs at all, and it's an important question, but I'd like to take a different angle and consider why there are so many visible bugs. Why does a bug in a driver, or even an application, bring down a whole system? In addition to reducing the incidence of actual bugs, IMO, we should also do a better job of containing the bugs that will inevitably exist even if we all use the latest whiz-bang code analysis tools (which rarely work for kernel code anyway). Some of the semi-informed members of the audience are probably thinking that's the job of the operating system; I'd argue that our entire current notion of operating systems is flawed. There are way too many components in a typical computer system that "trust each other with their lives" in the sense that if one dies all die. Memory protection between user processes is great, but there should be memory protection between kernel entities, and other kinds of protection, as well. One of the basic services that operating systems need to provide going forward is greater fault isolation and graceful instead of catastrophic degradation.
The Recovery Oriented Computing project at Berkeley has gotten some press recently for trying to address this issue. Many here on Slashdot don't seem to "get it" because they've never worked on systems in which a component failure was survivable; they don't realize that rebooting a single component - perhaps even preemptively - is better than having the whole system crash. "Software rot" is a real problem, no matter how hard we try to wish it away. ROC isn't about saying bugs are OK; it's about saying that bugs happen even though they're not OK, and let's do the best we can about that. Another project in the same space, with more of a hardware/security orientation, is Self Securing Devices at CMU. There, the idea is to find ways that parts of a system can work together without having to share each others' fate. While the focus of the work is on security, it shouldn't be hard to see how much of the same technology could be applied to protect a system from outright failure as well as compromise. There are plenty of other projects out there trying to address this problem, but those are two with which I happen to have personal experience.
The key idea in all cases is that current OS design forces us to put all of our eggs in one basket, and that's really not necessary. Designing fault-resilient systems is tough - few know that better than I do - but that's only a reason why we should do it once instead of devising ad-hoc clustering solutions for each specific application. Lots of people use various forms of clustering as a way to achieve fault containment and survive failures, but the solutions tend to be very ad-hoc and application-specific. Do you think Google's solution works for anything but Google, or that a database transaction monitor is useful for anything that's not a database? Fault containment needs to be a fundamental part of the OS, not something we layer on top of it.
Slashdot - News for Herds. Stuff that Splatters.
Microsoft says so.
Actually it's in some driver, not the core OS, so it's not surprising that it doesn't happen to everyone. (There's a few other things with similar problems.)